Online joint classification and anomaly detection via sparse coding

We present a novel convex scheme for simultaneous online fault classification and anomaly detection in a multivariate time-series setting. Our approach extends recent work on sparse coding and anomaly detection using an over-complete dictionary to problems where some taxonomy of anomalies already exists. The temporal aspect of the data is addressed by a simple sliding window approach; inspired by a group-LASSO penalisation approach, classification is treated by jointly sparsifying groups of the coefficients (the sparse coding) of dictionary atoms via ℓ_2;1 regularisation. The dictionary which drives the prediction and coding is assumed given and is learnable by a range of available prior algorithms. We demonstrate our framework on a classification and anomaly detection task on three-phase low-voltage time-series. In this case, we manually design our dictionary based on basic knowledge of common faults that affect low-voltage powerlines. For this reason our approach does not necessarily require a training stage.