In 2003, an electrical glitch in a power grid sparked one of the most extensive blackouts in history—the Northeast Blackout, affecting 50 million people across North America. This incident highlighted the need for predictive analytics to anticipate equipment malfunctions within such critical infrastructures.
Predicting the remaining useful life (RUL) of machinery and systems is vital across many sectors, including manufacturing and aviation, and involves gathering and analysing data to forecast future equipment failures using advanced machine learning (ML) techniques.
Yucheng Wang, a Senior Research Engineer at A*STAR’s Institute for Infocomm Research (I2R), noted that while newer ML methods like graph neural networks (GNNs) are being used to account for interactions among different sensor data, such as temperature and pressure, they often neglect local correlations. This oversight restricts how these models are built and ultimately limits their accuracy when predicting equipment lifespan.
To address these challenges, Wang and I2R colleagues worked with researchers from Nanyang Technological University, Singapore, to develop a new approach named LOGO (LOcal–GlObal correlation fusion). This framework integrates both immediate (local) and long-term (global) sensor data correlations into GNNs, enhancing the predictive accuracy. LOGO meticulously constructs models to represent sensor interactions from both perspectives, then uses these models to capture dependencies that evolve over time and across different sensors.
LOGO divides sensor data into smaller segments or 'patches', each processed to form sequential micro-graphs. Known as multi-patch segmentation, this action allows for the detailed analysis of local correlations, while global correlations are processed separately. An adaptive fusion mechanism then integrates these insights, ensuring each patch reflects a comprehensive spectrum of data.
The research team demonstrated the method's efficacy in numerous tests, which markedly outperformed traditional models and significantly reduced prediction errors to promise substantial cost savings and improved reliability. LOGO's success may enhance operational efficiency across various industries.
"This algorithm can be applied to aircraft engines to detect whether the engines need maintenance or repairs," Wang noted.
Looking ahead, the team aims to refine their process further. “The graph construction and GNN processes require a large number of samples for training. To address this and improve the model, we plan to incorporate data-efficient algorithms, such as self-supervised learning techniques,” Wang said.
The A*STAR-affiliated researchers contributing to this research are from the Institute for Infocomm Research (I2R).
