Enhancing Remaining Useful Life Prediction with Ensemble Multi-Term Fourier Graph Neural Networks

Abstract

Remaining useful life (RUL) prediction is crucial in predictive maintenance. Recently, deep learning forecasting methods, especially Spatio-Temporal Graph Neural Networks (ST-GNNs), have achieved remarkable performance in RUL prediction. Most existing ST-GNNs require searching for the graph structure before utilizing GNNs to learn spatial graph representation, and they necessitate a temporal model such as LSTM to leverage the temporal dependencies in a fixed lookback window. However, such an approach has several limitations. Firstly, it demands substantial computational resources to learn graph structures for the time series data. Secondly, independently learning spatial and temporal information disregards their inherent correlation, and thirdly, capturing information within a fixed lookback window ignores long-term dependencies across the entire time series. To mitigate the issues above, instead of treating the data within the lookback window as a sequence of graphs in ST-GNN methods, we regard it as a complete graph and employ a Fourier Graph Neural Network (FGN) to learn the spatiotemporal information within this graph in the frequency space. Additionally, we create training and test graphs with varying sizes of lookback windows, enabling the model to learn both short-term and long-term dependencies and provide multiple predictions for ensemble averaging. We also consider scenarios where sensor signals exhibit multiple operation conditions and design a sequence decomposition plugin to denoise input signals, aiming to enhance the performance of FGN. We evaluate the proposed model on two benchmark datasets, demonstrating its superior performance on the RUL prediction task compared to state-of-the-art approaches.