Real-time Outlier Detection in Time Series Data of Water Sensors

We compare multiple methods for real time outlier detection in time series data of water sensors. We present an outlier detection pipeline for this purpose. Multivariate models as well as univariate models are compared empirically by adding simulated outliers to the data to assess model performance. Quantile regression performed by the multi layer perceptron model using the tilted loss function is apt to model time series in a multivariate approach, provided we have access to reliable, correlated time series. Univariate models like auto-regressive models can be useful for detecting specific kinds of outliers such as extreme values. We show that the models are able to detect realistic, real life outliers.