Classification-based Diagnosis Using Synthetic Data from Uncertain Models

Abstract

Machine learning based diagnosis engines require large data sets for training. When experimental data is insucient, system models can be used to supplement the data. Such models are typically simplified and imprecise, hence with some degree of uncertainty. In this paper we show how to deal with uncertainty in synthetic training data. The data is produced using a model with uncertainties. The uncertainties originate from inaccurate parameter values or parameters that take dierent values based on the mode of operation. We demonstrate how techniques from the uncertainty quantification field can be used to reduce the numerical complexity of the training algorithm. In particular, we use generalize polynomial chaos to eciently approximate the loss function. In addition, we present a neural network architecture specifically designed to deal with uncertainties in the training data. As an illustrative example, we show how our approach can be used to detect faults in an elevator system.