Classifier Discriminant Postprocessing for Fault Understanding and Prognostics

The Multilayer Perceptron (MLP) approximates the Bayes optimal discriminant function and hence provides posterior probability estimates of classes it was trained upon. In the scope of Prognostics and Health Management (PHM), the MLP outputs an estimate of the likelihood that a given health class (e.g. system state or condition) is present in the input data. In this paper, we propose methodologies to use information embedded in the discriminant to not only provide system state classification, but also for: (1) change analysis, i.e. detecting short/long term changes in time-series data which allows to uncover incipient faults prior to full development; (2) new/unknown class detection, where the variance and sum of a discriminant across output neurons can be used to detect unknown class patterns in the input data; and (3) prognostics and remaining useful life estimation, where the fault-only part of the discriminant can be extracted and then fit to a nonlinear degradation model to estimate the evolution of a fault and the time that it becomes fully developed. Algorithms to extract these types of information may vary, and our methodology here is not constrained to any specific processing pipeline. The MLP may also be a part of a larger neural network architecture. For example, fully connected layers at the end of a deep Convolutional Neural Network (CNN) can be post-processed in a similar way. Thus, the applicability of our approach extends to both traditional and newer deep learning architectures.