Finding meaningful representations of SCADA-log information for data-driven condition monitoring applications

Analysis of data from supervisory control and data acquisition (SCADA) systems to monitor wind turbine
condition has attracted considerable research interest in recent years. Most approaches utilize time series
data from sensors placed all over the turbine and apply methods from the machine learning (ML) domain
for early failure detection (compare [1]). Besides, SCADA systems usually produce log-files which contain
information about operation conditions or control events but also warning and alarm messages in case
sensors measurements come close to or exceed pre-defined limits. In the condition monitoring context, this
data is also of high value and has been used to filter and annotate sensor time series [2], identify message-
patterns related to failures ([3], [4], [5]) and predict unplanned stoppages directly from the log-messages
([3]). However, most of these approaches use time-sequence or probability-based analysis rather than ML
methods. One reason is that a ML model usually requires a numeric vector representation of fixed size
as input. The challenge of finding such representations for symbol sequences of variable length is well
known in the ML domain, especially in natural language processing (NLP). Within this study, we extend
the Correlated Occurrence Analogue to Lexical Semantic (COALS) algorithm by incorporating temporal
information (COALS-t) and apply it to SCADA log-data. We demonstrate that the method can find
meaningful representations of SCADA log messages and message sequences. This enables the application
of off-the-shelve ML methods. The approach will be illustrated using multiple years of operational SCADA
data from several turbines. The remainder of the paper introduces the SCADA data set (section 2) and
the methods being applied (section 3) before presenting and discussing the results (sections 4 and 5).