PUB - Feature Generation and Evaluation for Data-Based Condition Monitoring of a Hydraulic Press

Content

Authors: Makansi, Faried; Schmitz, Katharina

Machine Learning approaches provide powerful means for capturing complex patterns from data. Therefore, their use in modern maintenance strategies is compelling. However, most algorithms are not suited to process raw machine data, but require a compression and transformation of the data to few characteristic features. While several code-toolboxes aim to facilitate feature generation by generic feature extraction, this contribution presents an extended procedure for refining and selecting automatically generated features. For the use case of a hydraulic press, the extended feature generation and selection proves to increase accuracy in fault classification while simultaneously providing insight into the effect of faults on measurable quantities of the system.