Remaining useful life prediction using graph convolutional attention networks with temporal convolution-aware nested residual connections
Yupeng Wei, Dazhong Wu, Janis Terpenny
Abstract
Degradation of engineered systems can result in poor performance and failure. In recent years, Graph Convolution Networks (GCN) have gained considerable attention for predicting the remaining useful life (RUL) by analyzing condition monitoring data. Conventional GCNs typically stack multiple spectral graph convolutional layers, where each layer aggregates condition monitoring data at different time and projects the aggregated data into another feature space. However, conventional GCNs suffer from two primary issues. Firstly, repeated aggregation operations in GCN layers can severely compromise the temporal correlation of condition monitoring data. Secondly, repeated aggregation and projection operations may generate less significant features, resulting in poor prediction performance. To address these issues, we introduce a temporal convolutional operation to extract and preserve temporal features before repeated aggregation and projection operations. Additionally, we incorporate an internal residual connection to skip some aggregation and projection operations to reduce the negative impact of less significant features. Finally, we use attention mechanisms after GCN layers to extract the most significant parts of features obtained from previous GCN layers and pass them to next GCN layers. We demonstrate the effectiveness and generalizability of our method through three case studies. Our numerical results show that our approach outperforms existing data-driven methods.