From riding in an elevator to booking movie tickets on your mobile phone, machines have become an integral part of life. But like all things on this planet, they have a set lifespan: despite working with unprecedented efficiency, even modern machines will eventually succumb to failure.
To avoid the consequences of a machine failing at random times, data-driven technology is used to predict a machine’s remaining useful life (RUL) so that machinery near the end of its lifespan can be pre-emptively maintained or replaced. These predictive algorithms are built around available source data on similar machines in similar working conditions that have been operated to failure.
Unfortunately, the accuracy of RUL prediction hinges heavily on how similar the working environments are for both the source and target machines. Furthermore, running a machine to failure just to gather training material can be impractical and costly. Mohamed Ragab of A*STAR’s Institute for Infocomm Research (I2R) devised a way to create a better predictive algorithm: one capable of maximum versatility and accuracy with minimal source data.
“Our idea is to identify the commonalities between the source and target domain, and to capture the unique characteristics of the target domain,” Ragab explained, adding that a more comprehensive representation of the target domain will boost an algorithm’s performance compared to existing approaches.
To this end, Ragab and his collaborators created the Contrastive Adversarial Domain Adaptation (CADA) algorithm.
CADA was built using training material from C-MAPSS, a public dataset that simulates run-to-fail experiments for engines under four different environments, all with varying operating conditions, fault types and lifespans. The team put CADA’s accuracy to the test using source data from one environment, with the remaining three acting as targets. The test was conducted on all four environments to create 12 source-target scenarios.
When pitted against conventional approaches, the researchers found that CADA was superior, with an accuracy of 38 percent compared to 21 percent using the traditional method. “CADA performs domain adaptation by producing source-like features, as well as preserving target-specific features by maximizing mutual information transfer,” said Ragab.
CADA’s versatility translates to greater accuracy and more flexibility in varying target environments. The algorithm can be trained on one source domain and applied to many target domains featuring unique working conditions.
Apart from potentially benefiting the manufacturing, semi-conductor, and healthcare industries, Ragab hopes to develop CADA to be trained and tested on the fly for even more real-world applications.
“This refining and updating can help reduce the processing time and improve its productivity and effectiveness,” he concluded.
The A*STAR-affiliated researchers contributing to this research are from the Institute for Infocomm Research (I2R).