Hierarchical Classification for Unknown Faults

Abstract

Data-driven prognostics and health management (PHM) models are generally trained on a set of data collected from the system under study. A standard assumption of this paradigm is that the training data contains all the normal operating conditions and fault conditions that are possible. If the training data does not contain all possible conditions, a single classifier approach will not be adequate because the PHM model could have difficulty classifying a new condition not previously seen during training. This study investigates the use of hierarchical classification in situations where the training data is incomplete in terms of the faults that are present in the testing set and characterizes the proposed problem as a transfer learning problem. The hierarchical classifier employs non-mandatory leaf node prediction where the model is not required to move to the lower levels of the hierarchy. It is hypothesized that this construction allows the classification to stop at a higher level when the fault is not present in the training data. The proposed method is demonstrated on a hydraulic actuator condition monitoring data set.