Segment-Based Graph Convolutional Network for Imputation of Missing Values Caused by Sensor Fault and Its Application in Aeroengines

Graph neural networks (GNNs) have emerged as a popular solution for missing value imputation. Nevertheless, traditional GNNs consume a significant amount of memory when processing long time series, limiting their deployment in resourceconstrained environments. Additionally, predefined static graphs are insufficient for modeling the dynamic nature of real-world time series. To address these issues, we propose a patch-based bidirectional graph convolution network (PBGCN) for imputing continuous missing values caused by sensor faults in long time series. Specifically, to alleviate the computational burden, we segment and embed long time series into several patches with fewer time steps, and model the temporal dependencies of these patches using the self-attention mechanism and the recurrent neural network. Simultaneously, we construct dynamic graphs that adaptively learn variable correlations based on data observed within a sliding window. A convolutional module built on these dynamic graphs is employed to facilitate patch-level information exchange between variables. Finally, the obtained temporal dependencies and variable correlations are processed in both forward and backward directions to achieve precise imputation. The results demonstrate the effectiveness of the proposed model. The proposed method matches or outperforms popular comparison methods for missing value imputation while reducing memory consumption by approximately 60% to 75% compared to traditional graph-based models.