Just as tyre treads wear out after years on the road, mechanical parts in manufacturing plants also suffer from the impact of wear and tear over time. Tool degradation is a problem in industrial settings because it can drive up manufacturing costs while diminishing product quality.
Machine learning (ML) offers a ray of hope for manufacturers seeking to optimise production lines. These platforms can learn from real-world data to predict the ‘sweet spot’ when manufacturing parts need to be serviced or replaced to maximise their lifespans without affecting product quality.
Amirabbas Bahador, a Product Development Scientist at A*STAR’s Advanced Remanufacturing and Technology Centre (ARTC), said that a specialised technique called transfer learning offers advantages over traditional ML methods for tool wear prediction.
“Transfer learning is done by starting from previously-learned patterns, instead of starting the learning process from scratch, saving significant amounts of manufacturing time and costs,” explained Bahador.
A group led by Bahador and Chunling Du, a Scientist at ARTC, combined transfer learning, low-cost sensors and one-dimensional convolutional neural networks to develop a first-of-its-kind tool wear prediction model.
In their study, the researchers tested their transfer learning system using two types of low-cost, microscale sensor-incorporated accelerometers: a microelectromechanical system (MEMS) and an integrated electronics piezoelectric (IEPE). These sensors detect linear motion, acceleration and shock in the machines they are attached to.
As the MEMS accelerometer was a single-axis accelerometer with limited capabilities, the transfer learning model significantly increased the tool wear detection accuracy using the MEMS accelerometer from 58 to 85 percent.
By leveraging knowledge gained from previously learned tasks, transfer learning was able to improve the accuracy of the tool wear detection model and reduce the amount of data required for model development. The team reported that their platform maintained high accuracy levels of 80 percent and above, even with up to 80 percent less training data.
Bahador said that developing the transfer learning model was no small feat: “The biggest challenge in designing the transfer learning model was selecting the number of layers required to be fixed or frozen from the source model, so that the transfer learning would have a high accuracy.”
The study’s success proves the utility of transfer learning for the manufacturing industry and even other sectors such as finance, marketing, and transport navigation systems. The team is continuing to push the limits of the technology for manufacturing applications.
“My next research focus would be to apply similar ML and transfer learning techniques to additive manufacturing process monitoring and powder characteristic evaluations,” concluded Bahador.
The A*STAR-affiliated researchers contributing to this research are from the Advanced Remanufacturing and Technology Centre (ARTC).