Imaging Wind Turbine Fault Signatures Based on Power Curve and Self-Organizing Map for Image-Based Fault Diagnosis

Wind turbine fault diagnosis based on time-series data from the supervisory control and data acquisition system, remains a challenging task. Due to various stochastic effects, one of the main challenges includes the fact that, a change in a few data points is not necessarily an indication of a fault; hence, how to interpret time-series trends is often a major challenge. In this paper, we propose an innovative approach which aims at encoding power curve time-series data into imagery data using the self-organizing map. This approach enables one to expose fault signatures in terms of data topological structure for a subsequent image-based fault diagnosis purpose. The obtained experimental results based on data from a real operational wind turbine, indicate the effectiveness and impact of the proposed method, when used as a primary procedure for an image-based fault diagnosis system.

I. Introduction

Condition monitoring of wind turbines (WTs) based on measurements from the supervisory control and data acquisition (SCADA) system, has received much attention in recent years thanks to the positive impact in terms of cost-effectiveness in wind farm’s operation and maintenance (O&M) activities [1 –4]. The impact is apparent both in industry and academia, where various research papers have extensively advocated as a cost-effective solution compared with other alternatives such as the so-called “standalone condition monitoring systems” (CMS) [5–7]. However, the use of SCADA data for WT condition monitoring purpose is not exempt of limitations. The main concerns are described as follows: (i) Wind turbine SCADA data are generally recorded with an interval of 10 minutes average. Since the signals are recorded with a long interval, which initially was not designed for the purpose of condition monitoring, most dynamic features of WT faults that are useful for condition monitoring are lost. As a result, the detailed information, for instance the location and mode of most WT faults cannot be diagnosed by using the conventional signal processing techniques such as time-domain, frequency-domain (spectral analysis), and time-frequency domain (wavelet transform) analyses [8, 9]. Instead, methods based on SCADA data usually resort to “mathematical statistics” or “intelligent learning” methods, in order to effectively extract features from the historical or current observations [5]. To date, there is no standard approach. (ii) The SCADA data of WTs are collected under varying operating conditions. In addition to the nonlinear control effects that could damper the features of faults, it is still difficult to detect an incipient fault from SCADA data, unless advanced data analysis tools are developed or the fault is severe. (iii) In general, time-series-based diagnosis methods for WTs face a major challenge in interpreting trends whose deviations in a few data points is not necessarily an indication of a fault [7].