A few-shot based foundational fault detection model based on gear test bench data

Few-shot learning reduces the data required for vibration-based condition monitoring. However, current methods struggle to generalize to new machine and fault instances. This limits industrial applicability, where diagnosing developing faults is crucial and fleets of machines are common. We propose a few-shot learning-based fault detection method that can be trained on separate datasets and then used on a target machine from which no fault samples are available. Early healthy samples from the target machine are utilized to establish a baseline healthy state. The method shows strong generalization when validated on multiple run-to-failure gear experiments on a large test rig