Multistage Graph Convolutional Network With Spatial Attention for Multivariate Time Series Imputation | JoVE Visualize

Area of Science:

Data Science
Machine Learning
Time Series Analysis

Background:

Data loss in multivariate time series (MTS) analysis degrades model performance for critical applications like structural health monitoring (SHM).
Existing MTS imputation methods fail to account for heterogeneous correlations arising from diverse sensor types and dynamic environmental conditions.
Accurate imputation of missing MTS data, especially considering heterogeneous and dynamic correlations, remains a significant challenge.

Purpose of the Study:

To develop an advanced imputation method for multivariate time series (MTS) that effectively handles heterogeneous and dynamic correlations.
To improve the accuracy of data imputation in real-world applications, such as structural health monitoring and traffic flow monitoring.
To propose a novel deep learning architecture capable of learning complex inter-variate relationships in heterogeneous MTS data.

Main Methods:

Proposed a multistage graph convolutional network with spatial attention (MSA-GCN) for MTS imputation.
Stage 1: Decomposed heterogeneous MTS into homogeneous clusters to learn intracluster correlations.
Stage 2: Employed a graph convolutional network (GCN) with spatial attention to capture dynamic intercluster correlations.
Stage 3: Utilized stacked convolutional neural networks for feature decoding and missing data prediction.

Main Results:

MSA-GCN demonstrated superior imputation performance compared to baseline models on diverse datasets.
The method effectively learned heterogeneous and dynamic correlations inherent in MTS data.
Significantly reduced imputation errors in real-world datasets, including bridge monitoring and weather data.

Conclusions:

MSA-GCN offers a robust solution for accurate MTS data imputation by effectively modeling complex correlations.
The proposed architecture addresses the limitations of existing methods in handling heterogeneous and dynamic sensor data.
The findings highlight the potential of MSA-GCN for enhancing downstream tasks reliant on complete and accurate MTS data.