Multivariate Time-series Data Correction by combining Attention-based LSTM and GAN Model

Abstract

High-quality data can increase the reliability of machine learning-based prediction models. In our work, we propose a novel method for data correction to improve the quality of multivariate time-series data. For this, we use a LSTM-based VAE-GAN for anomaly detection and an Attention-based LSTM model for data correction. Through experiments using Secure Water Treatment (SWaT) data, we show that the proposed correction method is superior to previous correction methods.