Statistically enhanced classification of centrifugal compressor operating condition

Centrifugal compressors operating near peak efficiency are prone to aerodynamic instabilities, which in extreme cases, may lead to quick destruction of the machine. Instabilities are flow structure of different character, which are not straightforward to detect. One of the possible approaches could take advantage of feature space representation. Historic data might be used for building clusters of operating conditions and defining class boundaries. Then, a new point could be classified based on its location in feature space relative to those boundaries. Different instabilities can coexist temporarily when transitioning from conditions to another, which makes defining the boundaries in the transition zones a challenge. To deal with those problems, an application of Gaussian Process Classification (GPC) method is proposed. Being a nonparametric method of high flexibility, it can provide better separation between conditions in the feature space. GPC output provides a confidence level of a new observation belonging to a class, which can be used for classification but also to introduce a no-classification zones based on class probability values, which could decrease the error rate in transition zones. In this study, GPC is applied to classify features obtained from physically interpretable feature extraction method for a centrifugal compressor. Three different classes are defined, one stable and two related to different instabilities. The classification with GPC is compared to threshold approach, based on physical interpretation of the features. The results demonstrate that GPC offers a more flexible and data driven separation of classes, especially in transition zone between different flow conditions, however the rejection zone does not prove to provide improvement in the analysed case. Compared to the threshold approach, GPC requires historic data for all conditions and results depend strongly on data labels. If the physical interpretation of the features is possible, the advantages of GPC in the transition zone can be combined with the binary threshold approach in well-defined feature space regions.