Reconstruction-based time series anomaly detection on contaminated data

Time series anomaly detection(TSAD) aims to find abnormal segments in sequences. There have been many methods are proposed for time series anomaly detection, including reconstruction-based and prediction-based methods. Previous reconstruction-based time series anomaly detection models are implicitly trained to perform poorly on anomalies, so that anomalies will stand out in residuals. But some unexpected fitting issues may allow the models to still reconstruct anomalies well in time series. For example, the default no-anomaly assumption of the training set in previous researches may make the models fit the hidden anomalies, causing a degraded detection sensitivity. Toward a more rigorous detection, we innovatively introduce the Drop-Loss to unsupervised eliminate possible anomalous losses during training to block the backpropagation of abnormal gradients, thereby avoiding unexpected fitting. The proposed method is validated on four benchmarks with extensive experiments and achieve state-of-the-art results.