EdgeCog: A Real-Time Bearing Fault Diagnosis System Based on Lightweight Edge Computing

Application of the Nadaraya-Watson estimator based attention mechanism to the field of predictive maintenance

“…Previous work done in this area focused mainly on complex deep learning techniques [ 12 ] (see Section 6 ), while we devise a method for use of simpler and incredibly fast techniques that are less demanding on compute resources. Our work, inspired by [ 3 ] and [ 13 ], applies the Nadaraya-Watson kernel as a simple but effective attention mechanism to predict tool-wear for a milling machine that undergoes non-linear wear characteristics shown in Fig.…”

“…3 . When compared to deep learning based techniques such as [ 12 ] and [ 16 ], both SOTA and recent (2023), our method is much simpler to program, train and use and validated on a benchmark milling tool-wear use-case. Deep learning methods are necessary for fields such as NLP, as features to the order of millions or even billions (i.e.…”

“…In the maintenance use case, the matrices are much smaller (order of 103 rows and less than 10 columns), where simpler ML techniques may prove to be equally effective. Deep learning methods create large models and usually need quantization (reducing floating point precision) to reduce its size, [ 12 ]. Our method is natively lighter and fast, and therefore suitable for deployment on edge computing devices.…”

“…When compared to deep learning based techniques such as [ 12 ] and [ 16 ], both SOTA and recent (2023), our method is much simpler to program, train and use and validated on a benchmark milling tool-wear use-case.…”

“…Deep learning methods create large models and usually need quantization (reducing floating point precision) to reduce its size, [ 12 ]. Our method is natively lighter and fast, and therefore suitable for deployment on edge computing devices.…”