The objective of this article is real-time detection of a change-point in the baseline distribution of the frequency signal generated by Phasor Measurement Units (PMUs) that could indicate potential for voltage collapse, false data injection, or other security threats. The alarm flags an anomaly event when the cumulative sum (CUSUM) of the log-likelihood departures of the Hurst exponent of the frequency from its baseline statistic exceeds a threshold set up as a compromise between the conflicting objectives of minimum detection delay and acceptable False Alarm Rate. As main theoretical contribution, an extra protection layer is developed that provides an estimate of the time to the alarm event, giving the Transmission System Operator (TSO) or an autonomous agent more time to deploy proactive measures to avoid large catastrophic system states. The proposed method is illustrated by a retrospective analysis of the 2012 Indian blackout that reveals that 10-12 minutes before the voltage collapse, a significant increase in the Hurst exponent of the frequency could have been detected. Conceptually, this shows that PMUs provide significant value towards ensuring autonomous cognition in the power grid by enabling the abnormal events to be fault-detected in an unsupervised fashion.

We introduce a novel optimization strategy that combines empirical risk minimization (ERM) with a sample weighting scheme to decrease the bias induced by training with noisy labels. We accomplish this by penalizing the tails of the loss distribution, as samples associated with the right tail (high loss values) are more likely to be mislabeled. We implement the proposed sample weighting scheme by minimizing a discriminatory risk that downweighs mislabeled samples. Moreover, we show that the proposed method enables us to optimize the location and shape of the loss distribution simultaneously. In addition, we couple our method with a distributionally robust optimization stage. The goal of this coupling is to increase the weight of correctly labeled but hard-to-Iearn samples that exhibit a high loss value. This coupling also improves model accuracy on underrepresented minority subsets when training with imbalanced classes. Experimental results show that for 60% noise contamination, our method, when compared to the ERM, improves the final accuracy on both Fashion-MNIST and CIFAR-10 by 18.4% and 8.5%, respectively.