One Class Data Mining Approaches for Leakage Fault Detection in Refrigeration Showcases

Abstract

In the area of fault analysis, it is usually the case when labeled data is not available for training accurate classifiers in discriminating between the normal function conditions and abnormal events. This paper will address the problem of fault identification for refrigeration showcases, which also follows the aforementioned setting, and whose continuous and ample use in commerce requires that unusual conditions and faults to be identified as quickly (and reliably) as possible. While the data for one class is available in problems of this nature, we will be considering not only specific algorithms for unary classification, but also the performance in the application of supervised and unsupervised machine learning algorithms. Results showed the value of the implemented approaches in narrowing down anomalous events with more consistency, when compared to the statistical method standardly employed to this task.