CERN, Meyrin, Switzerland
Universität Stuttgart, Stuttgart, Germany
Machine learning algorithms enable failure prediction of large-scale, distributed systems using historical time-series datasets. Although unsupervised learning algorithms represent a possibility to detect an evolving variety of anomalies, they do not provide links between detected data events and system failures. Additional system knowledge is required for machine learning algorithms to determine the nature of detected anomalies, which may represent either healthy system behavior or failure precursors. However, knowledge on failure behavior is expensive to obtain and might only be available upon pre-selection of anomalous system states using unsupervised algorithms. Moreover, system knowledge obtained from evaluation of system states needs to be appropriately provided to the algorithms to enable performance improvements. In this paper, we will present an approach to efficiently configure the integration of system knowledge into unsupervised anomaly detection algorithms for failure prediction. The methodology is based on simulations of failure modes of electronic circuits. Triggering system failures based on synthetically generated failure behaviors enables analysis of the detectability of failures and generation of different types of datasets containing system knowledge. In this way, the requirements for type and extend of system knowledge from different sources can be determined, and suitable algorithms allowing the integration of additional data can be identified.
