Industrial Pumps Anomaly Detection and Semi-supervised Anomalies Labeling Through a Cascaded Clustering Approach

Abstract

Automation technology has brought significant changes to agriculture, industry, commerce and other fields, among which the machine learning algorithms are the important applications of predictive maintenance of industrial equipment. In general, anomalous trends should be detected timely before failure occurs so that unscheduled downtime can be avoided. In addition, predictive maintenance can avoid unnecessary maintenance and make good use of component remaining life by setting appropriate maintenance periods for worn parts. In this paper, based on the real case in which data collected by the various sensors on coal mine pumping system, we propose a cascaded unsupervised clustering method that consists of DBSCAN and spectral clustering to identify uncommon abnormal data and classify the common abnormal data. As equipment continuously operating, the proposed cascaded clustering method can gradually utilize the obtained uncommon abnormal data to enlarge the common abnormal data. This process implemented through periodic manually labeling is regarded as a semi-supervised manner. The results show that DBSCAN has good discriminative power for uncommon abnormal data, and the spectral clustering can properly classify working condition of water pumps with 93% accuracy on test data.