Datasets of historical performance metrics can offer valuable insight into an asset fleet’s health. This is especially so in the context to establishing normal behavior and thresholds of acceptable performance for diagnostic purposes. However, plant performance can often be obscured by data quality issues which introduce artefacts that do not pertain to asset health. This paper utilises a supervised ensemble machine-learning approach to automate the process of filtering maintenance data based on their predicted validity. The results are then presented both in terms of classification performance, and the impact on the distributions directly. This helps to ensure engineers are basing their diagnostic decisions on valid data. The accuracy of the filtration process, and its effect on the final thresholds will be discussed. To illustrate, this paper uses data of varying quality on circuit breaker trip tests obtained from operational medium-voltage circuit-breakers spanning several decades with the aim of providing decision support for switchgear diagnostics.
|Number of pages||6|
|Journal||PHM Society European Conference|
|Publication status||Published - 1 Jul 2018|
|Event||Fourth European Conference of the PHM Society - Utrecht, Netherlands|
Duration: 3 Jul 2018 → 6 Jul 2018
- machine learning
- circuit breakers