Prognostics of complex machinery with sparse multilabel multimodal run-to-failure data: A graph neural network approach

Abstract

The practical requirements for maintaining machine operability often conflict with the data needs for training prognostics models due to the limited availability of run-to-failure (RTF) data in industry settings. This scarcity is exacerbated by irregular and infrequent inspections, resulting in sparse datasets. The literature tends to address this challenge by trajectory data augmentation methods which creates more RTF data by transformations on available trajectories, but these methods still require some trajectories to begin with. To address the challenge faced by industries where running any machine to failure without intervention is impractical, we propose a diagnostics feature similarity-based method to construct full RTF trajectories from partial data, which is then used in a graph neural network for prognostics. Unlike conventional graph-based prognostics that primarily model sensor interactions through static graph structures, this research explores fault propagation as an evolving graph, a novel approach in the application of GNNs. It posits that condition monitoring data from various machines across diverse health states can effectively generate prognostic insights and model degradation evolution as a dynamic graph with physically meaningful node-edge embeddings. The efficacy of this method is demonstrated through its application in a hydrogenerator prognostics case study involving multiple fault states.