Learning to Detect Industrial Time-Series Anomalies From Imputation Consistency With Sparse Observations

Time-series anomaly detection plays an important role in ensuring industrial safety. Currently, many anomaly detection methods mainly target complete time series and ignore the widespread problem of data missing in the real world. Therefore, this article proposes a novel anomaly detection method for time series with sparse observations based on imputation consistency using a mixture of patch information inference network (MoPIN). Due to the robustness of the imputation method modeling to the random mask, different imputed series of the same normal time series with different random masks should have consistency. Then, a novel imputation consistency is used to detect anomalies in sparse observation series. Moreover, the MoPIN imputes series by a two-step imputation and multiscale modeling of patch information. Meanwhile, the similarity of imputed series under different masks is used to measure imputation consistency, which well constructs the relationship between sparse observation series and anomaly scores. Finally, the MoPIN can accurately detect anomalies while imputing series. Extensive experiments on four real-world benchmarks in different domains of imputation and anomaly detection tasks and a real fluid catalytic cracking (FCC) process case demonstrate the effectiveness of the proposed method. Specifically, the MoPIN achieved at least 8.05% mean absolute error (MAE) relative improvement in imputation and 3.74% $F1$ relative improvement in anomaly detection.