Experimental investigation of the effectiveness of K-means for classifying misalignment and unbalance faults in industrial rotating machines

The automation of condition monitoring techniques especially vibration analysis is increasingly emphasized amongst academic and industrial reliability experts, owing to the challenges posed by big data management as well as the need to reduce subjective judgements that may arise from human interference. While existing body of literature indicates that researchers are investing signifi-cant efforts towards achieving this objective through the application of techniques such as artificial neural networks (ANN), support vector machines (SVM), fuzzy c-means, etc. Despite the appreciable contributions from these techniques, their practicability under real life industrial scenarios is still questionable due to the re-quirements for training both healthy and faulty data sets relating to different ma-chine conditions. Additionally, most of the previous studies have been conducted on bearings and gears with very few studies exploring the automatic classification of other commonly encountered rotating machine faults such as shaft misalignment and unbalance. Based on this premise, the current study investigates the effectiveness of a popular unsupervised pattern classification tool, K-means. The rationale behind selecting this particular tool (i.e. K-means) is primarily based on its computational simplicity as this could enhance industrial acceptability in the near future. Besides faults detection, condition monitoring also requires the optimisation of the diagnosis features so as to eliminate redundant features that may reduce diagnosis accuracy. The preliminary results obtained from this study suggest that a single frequency domain feature is capable of differentiating misalignment and unbalance at two machine speeds.