Unsupervised-ensemble-based method for automatic running-in information extraction in reciprocating compressors

This work presents a fully automatic method for extracting running-in information from data of hermetic reciprocating compressors by analyzing clusters of subsequenced time series data. We used the $k$-means, kernel $k$-means, employing both a radial basis function and a novel application of the Mahalanobis radial basis function kernel, and agglomerative hierarchical clustering algorithms for clustering the data. The method is based on an ensemble of single occurrence transition detection models trained considering several parameter combinations and clustering algorithms. We developed a pruning method to identify the most meaningful transitions, discarding models whose results did not relate to the running-in process and allowing for feature interpretation based on the parameters of the remaining models. Experimental evaluation of the proposed method revealed that the electric current of the compressor is the most significant feature for tribological steady state detection and that the Mahalanobis-RBF kernel provides the best results. As a result, the proposed method offers an automated analysis of the running-in duration in hermetic compressors, potentially improving the reliability of compressor tests and saving resources in the preparation process.