Change-point detection in wind turbine SCADA data for robust condition monitoring with normal behaviour models

Abstract

Abstract. Analysis of data from wind turbine supervisory control and data acquisition (SCADA) systems has attracted considerable research interest in recent years. The data is predominantly used to gain insights into turbine condition without the need for additional sensing equipment. Most successful approaches apply semi-supervised anomaly detection methods, also called normal behaivour models, that use clean training data sets to establish healthy component baseline models. However, one of the major challenges when working with wind turbine SCADA data in practice is the presence of systematic changes in signal behaviour induced by malfunctions or maintenance actions. Even though this problem is well described in literature it has not been systematically addressed so far. This contribution is the first to comprehensively analyse the presence of change-points in wind turbine SCADA signals and introduce an algorithm for their automated detection. 600 signals from 33 turbines are analysed over an operational period of more than two years. During this time one third of the signals are affected by change-points. Kernel change-point detection methods have shown promising results in similar settings but their performance strongly depends on the choice of several hyperparameters. This contribution presents a comprehensive comparison between different kernels as well as kernel-bandwidth and regularisation-penalty selection heuristics. Moreover, an appropriate data pre-processing procedure is introduced. The results show that the combination of Laplace kernels with a newly introduced bandwidth and penalty selection heuristic robustly outperforms existing methods. In a signal validation setting more than 90 % of the signals were classified correctly regarding the presence or absence of change-points, resulting in a F1-score of 0.86. For a change-point-free sequence selection the most severe 60 % of all CPs could be automatically removed with a precision of more than 0.96 and therefore without a significant loss of training data. These results indicate that the algorithm can be a meaningful step towards automated SCADA data pre-processing which is key for data driven methods to reach their full potential. The algorithm is open source and its implementation in Python publicly available.

Detection of operating mode changes, without a priori model and in uncertain environments

“…Some applications are to be noted in the industrial field (Hawkins et al, 2003; Paul, 1994), energy consumption (Touzani et al, 2019), turbines (Letzgus, 2020), manufacturing processes (Atashgar and Rafiee, 2020; Guo et al, 2019; Wang et al, 2014), and chemical process (Salvador et al, 2014). We can also note the recent development of this theme in current environmental fields (Awe and Adepoju, 2020; Burnett, 2019) and societal fields (Choe et al, 2016; Coughlin et al, 2021; Salmasnia et al, 2020; Sürmeli et al, 2017).…”