Speaker
Description
Predicting equipment degradation and failure within the International Monitoring System (IMS) would enable more efficient maintenance intervals and ultimately reduce station downtime. The coarse sampling of the time-series data, the scarcity of known, labeled failures, as well as the potential influence of unknown failures in the training data makes an AI/ML solution challenging. We introduce a methodology that leverages several state of the art methods in unsupervised training, data augmentation, and model design to overcome these difficulties. We outline our training procedure and offer a series of diagnostic tests that can equally apply to time-series data sets with similar qualities. We show that this method can produce embeddings from time-series data that can be used to compute a system health score. We also illustrate a CONOPS where an algorithm trained using this method can learn to classify early signs of degradation from a database of observed failures. Finally, we discuss the key challenges of deploying such an algorithm and its potential impact on operations.
| [email protected] |
