DT-GAIN: a Novel Framework for Multivariate Time-Series Data Imputation in Industrial Iot
Kamran Sattar Awaisi, Qiang Ye, Srinivas Sampalli · 2025
In Industrial Internet of Things (IIoT) applications, data from a variety of sensors is collected continuously to monitor and manage industrial processes. However, missing data caused by network disruptions, sensor malfunctions, or hardware failures seriously affects the performance of datadriven models, leading to unreliable predictions and increased maintenance costs. To address this challenge, we propose Decayaware Transformer-enhanced GAIN (DT-GAIN), a novel imputation framework for multivariate time-series data in IIoT applications. DT-GAIN extends and enhances the Generative Adversarial Imputation Nets (GAIN) framework by integrating the Transformer architecture, which captures long-range dependencies essential for accurate imputation of multivariate timeseries industrial data. In addition, DT-GAIN incorporates a timedecay mechanism that accounts for temporal irregularities by weighing observations based on their recency, thereby improving the ability of the proposed method to handle varying intervals between observations and missing values. In our research, we thoroughly compare DT-GAIN with state-of-the-art imputation methods, including LSTM-based GAIN (L-GAIN), Transformerbased GAIN (T-GAIN), the original GAIN, and SAITS. Our experimental results indicate that DT-GAIN outperforms the methods under investigation in terms of Root Mean Squared Error (RMSE) across various missing rates, particularly excelling in high-missing-data scenarios.