Spatial-temporal multi-sensor information fusion network with prior knowledge embedding for equipment remaining useful life prediction

As the multi-sensor signals have the intricate spatial-temporal correlations and inherent heterogeneities, the current remaining useful life prediction methods based on graph neural networks (GNNs) are difficult to capture dynamic temporal patterns within each sensor's signal, and the predefined graph structures always fail to represent the evolving relationships between sensors over time. Moreover, the existing dynamic graph methods struggle to model long-term dependencies and effectively incorporate available priori knowledge. To address the above issues, a novel spatial-temporal multi-sensor information fusion network (SMIFN) is proposed. It primarily consists of a temporal information mining module (TIMM) and a spatial relationship modeling module (SRMM). Specifically, a bi-autoregressive attention mechanism is proposed in TIMM to extract the intra-sequence relationship within each sensor signal, meantime a self-attention mechanism is applied to mine the inter-sequence relationships of multi-sensor signals. These temporal features can be considered as deep node feature representations, and they are fed into SRMM for facilitating the fusion of spatial-temporal information. The proposed SRMM comprises graph construction and a multi-channel graph convolutional network, which can effectively capture the spatial dependencies among various sensors. Experiments on CMAPSS dataset and a real-world wind turbine SCADA dataset demonstrate that the proposed SMIFN has a better RUL prediction performance than the classical and state-of-the-art RUL prediction methods.