Smoothed LSTM-AE: A spatio-temporal deep model for multiple time-series missing imputation

Abstract

Missing values widely exist in time-series data owing to sensor or communication failure. It is indispensable to impute the missing data for equipment state monitoring and advanced data analysis. In this study, we propose a deep spatiotemporal time-series missing data imputation model, called LSTM-AEs, to enhance the imputation performance and handle multiple missing patterns. Generally, it is practically challenging to determine when, which, and how many sensors fail. Most previous algorithms cannot handle multiple-sensor missing data owing to insufficient attributes. As different missing patterns require different models and may appear simultaneously, it is inadvisable to employ a large number of models for handling all these patterns. A deep auto-encoder (DAE) can eliminate the diversity of missing patterns by restoring the information from the sensor data. The proposed model combines DAE and LSTM for extracting spatio-temporal features to estimate missing values in multiple time series. Moreover, a smoothing regularization term is added into the combined model, leading to a more stable estimation. Experiments are conducted on three types of sensor data: the Tennessee Eastman process simulation data, gas turbine data from the offshore oil Corporation, and power plant simulation data. The results demonstrate that the proposed technique is effective for different missing patterns and provides more accurate predictions than the existing techniques.