Data Modeling for Fault Detection

Data Models are high order [O(3)] multi-variable characterizations derived from simple third order polynomial building blocks. Such functions can be used to build real-time operational condition diagnostic and prognostic, i.e. fault classification and time to failure estimates. Data Modeling can achieve anomalous detection while only requiring nominal (no-fault) conditions for training. This makes Data Modeling an attractive tool for Novelty Detection and ambiguity resolution between nominal/anomalous; also Data Modeling can resolve ambiguities in diagnostic calls and manage risk uncertainty in prognosis estimates of time to fail. This paper presents the theory of how Data Modeling was successfully applied to Novelty Detection, and how it can be applied to Diagnostic Ambiguity Resolution and Prognostic Risk Uncertainty Management. Classifier methods such as adaptive multi-dimensional distance measure neural networks and Divergence classifiers were unsuccessful when applied to a set of diagnostic vibration feature vectors with only slightly anomalous conditions. When Data Modeling was applied to the same feature vector sets, 100% correct classification was achieved.