Dimensionality analysis in machine learning failure detection models. A case study with LNG compressors

Abstract

The correct selection of the subset of features for the design of a CBM (Condition Based Maintenance) strategy may result in models working faster and producing more accurate predictions. This must be done avoiding a phenomenon known as the curse of dimensionality, that appears in Machine Learning when algorithms must learn from an ample feature volume with abundant values within each one. This paper deals precisely with feature selection problem when dealing with compressors failure modes detection, using machine learning (ML) models. To that end, several feature selection ranking (FSR) meth- ods are considered. These methods are basically algorithms which include wrappers and filters and they are able to provide a ranking about all the analysed features. A very important issue of these methods, is to realise the feature selection unconstrained of the Machine Learning algorithm to be later applied, and that will be tested in this paper. Stability and scalability of these methods will be also defined and discussed in the paper. The paper case study evaluates the possibility of detecting and therefore diagnosing the rod drop failure mode appearance in Liquid Natural Gas (LNG) cryogenic reciprocating compressors by using artificial intelligence analysis techniques. This failure mode implies unavailability of this equipment, which are critic in the LNG industry due to the cost of flaring or, less common and desirable, venting of the boil-off gas (BOG) recovered by its compression in order to send it out or use as fuel. More than 90.000 running hours and thirteen representative features are evaluated as well as thirteen FSR methods. Three most-used classifiers have been employed in order to assess the feature rankers’ effect over the models development to diagnose the rod drop failure. Conclusions are about the possibility, not only to diagnose the appearance of a failure mode like rod drop, but also to do it with considering a reduced number of features.