Real-time streaming data takes on distinct visible patterns, known as regimes, as a result of changing external influences. Regimes corresponding to hazardous states, such as turbulent flow in oil pipelines or patients experiencing heart arrhythmias, must be identified quickly and accurately by on-line detection algorithms. In this paper, we propose a modification to the mixture of experts framework, which is traditionally used to model piecewise stationary time series. Our proposed modification allows experts to produce features specific to their designated regimes, rather than being limited to prediction error. This approach provides the flexibility to update the mixture modularly as new regimes emerge without the burden of retraining the entire mixture, as is typical in traditional classifiers. Our approach is tested on flow rate data from an oil and gas application, as well as detecting heart arrhythmias from electrocardiogram (ECG) signals. It outperforms traditional classification approaches both in terms of error rate and detector delay.
展开▼
