Just as all of us are advised to undergo regular health check-ups, machines are no different. To keep endless rows of factory machines in tip-top condition, a valuable parameter that can be tracked is a machine’s remaining useful life (RUL), which predicts the health of a machine component or system.
With accurate RUL prediction, businesses can reduce maintenance costs by performing maintenance only when necessary, and improve system reliability by minimizing machine downtime and maximizing machine lifetime. However, it is challenging to model these systems accurately, as different mechanical systems have their distinct mechanisms.
To boost the performance of RUL prediction, a research team that included Xiaoli Li, a Principal Scientist at A*STAR’s Institute for Infocomm Research (I2R), built a new RUL prediction framework that combines learned features derived from deep-learning algorithms, with handcrafted features derived from manual inputs by a data scientist.
“Both deep learning and handcrafted features with domain knowledge are important for accurate RUL prediction,” Li explained. “However, the effective combination of these two types of features is non-trivial.”
Their strategy, which they call an ‘attention-based deep learning framework,’ combines the best of both learned and handcrafted features: First, a ‘long short-term memory (LSTM) network’ is used to learn characteristic features in sequence; second, an ‘attention network’ automatically assigns larger weights to more important features; and third, a ‘feature fusion framework’ combines LSTM with handcrafted features.
In two real-world datasets, the researchers applied their attention-based deep learning framework to describe the deterioration of an aircraft engine. Based on two widely used criteria for evaluating RUL prediction performance, they showed that their approach outperformed current ‘state-of-the-art’ methods based on artificial intelligence (AI) algorithms.
“People often say deep learning can achieve the best prediction results,” said Li. “Our research demonstrates that human domain knowledge is equally valuable and critical for boosting system performance. By integrating AI and human intelligence, we can achieve the best possible prediction outcomes.”
In real-world scenarios where each machine operates under different working conditions, models trained under specific conditions may not generalize well to other machines. As such, the researchers plan to incorporate transfer learning in their RUL predictions, to overcome an assumption in machine learning that training data and test data share the same basic patterns and distributions. By developing an AI model that can self-adapt to new working environments, it will be possible to develop RUL predictions that more accurately predict complex interactions found in the real world.
The A*STAR-affiliated researchers contributing to this research are from the Institute for Infocomm Research (I2R).