Peering into the future like a digital crystal ball, artificial intelligence (AI) is transforming the way industries operate. Specialised algorithms can predict when a machine will require maintenance, or its remaining useful life (RUL), thereby reducing operational downtime and saving costs.
Ruibing Jin and Zhenghua Chen, Scientists at A*STAR’s Institute for Infocomm Research (I2R), are at the helm of AI-driven methodologies to forecast RUL with greater precision, pushing the boundaries of what’s possible in industrial efficiency and proactive maintenance.
“After reviewing existing approaches for RUL prediction, we find that these methods show unstable performances in different subsets which include different conditions and fault models,” said Jin.
Jin explained that traditional static neural networks typically have fixed architectures; these networks may not always ‘learn’ effectively from the changing behaviours of machines, hindering their ability to provide the most accurate and up-to-date information on a machine’s health and lifespan.
To navigate this challenge, Jin and Chen introduced a first-of-its-kind adaptive and dynamic neural network called AdaNet. This approach was designed to better capture sequential information in time series data and keep pace with evolving patterns of machine behaviour, thereby improving RUL prediction accuracy.
“To the best of our knowledge, we are the first to propose a dynamic network for RUL prediction,” Jin remarked.
The team ventured into other research domains to find a solution for adaptive network architectures. Their efforts culminated in the successful adaptation of deformable convolution techniques, typically used for visual data, to sequential time series data.
AdaNet distinguishes itself by altering its neural network kernels and selectively activating channels based on the incoming data, significantly honing its data extraction capabilities. When benchmarked against the C-MAPSS dataset, AdaNet outclassed other models in all six conditions, showcasing its superior early prediction capabilities.
"An early prediction of RUL allows industries to replace parts in a timely manner, circumventing industrial failures," Jin explained.
While AdaNet's immediate implications for RUL prediction are profound, its potential spans across other applications involving time series data, such as monitoring human activities. Currently, the research team is engaging with industrial partners to explore practical deployments of AdaNet, signalling a new horizon in the application of AI for industrial prognostics.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR’s Institute for Infocomm Research (I2R).
