Interpretable detection of unstable smart TV usage from power state logs

Power state logs from smart TVs are collected in order to construct a time-series representation of their usage. Time-series that belong to a TV exhibiting instability problems are classified accordingly. To do so, an automated feature extraction approach is used, together with linear classification methods in order to realise interpretable classification decisions. A normalized true positive rate of 0.84 ± 0.10 is obtained for the classification. The normalized true negative rate equals 0.80 ± 0.03. The final model returns a regularity statistic called the Approximate Entropy as its most important feature.