Learning to Classify Sensor Data

Learning how to classify sensor data is one of the basic learning tasks in engineering. Data from sensors are usually made available over time and are classiied according to the behavior they exhibit in speciic time intervals. I address the problem of classifying nite, univariate, time series that are governed by unknown deterministic processes contaminated by noise. Time series in the same class are allowed to follow diierent processes. In this context, I investigate the appropriateness of using induction algorithms not speciically designed for temporal data. I present a simple supervised induction algorithm that uses serial correlation as its inductive bias in a Bayesian framework, and compare it empirically to a popular general-purpose classiier, in a NASA telemetry monitoring application. Two comparisons were performed: one in which the general purpose classiier was applied directly to the data, and another in which features that captured serial correlations were extracted before the induction. Serial correlation appeared to be an important form of inductive bias, most eeectively utilized as an integral part of the learning algorithm. Feature extraction occurs too early in the training process to utilize correlation knowledge eeectively.