One-fault-shot learning for fault severity estimation of gears that addresses differences between simulation and experimental signals and transfer function effects

Gearboxes are integral elements in rotating machines and have a high tendency to fail due to their operation in harsh conditions. A robust method to estimate the fault size of gears is desirable for a successful prognostic process, which is, to date, still unavailable in the literature. The fault size can be estimated by vibration analysis, using signal processing and machine-learning tools. However, the availability of labeled or unlabeled vibration signals from faulty rotating machinery components is rare, making it challenging to apply machine-learning algorithms. Therefore, some physical pre-knowledge should be incorporated in the model for a successful learning process. This can be done by exposing the learning model to simulated data, and by a physical pre-processing procedure. This paper suggests a novel algorithm to overcome the lack of faulty data (labeled and unlabeled), and it is trained on a combination of simulated data and some real data. The algorithm tunes the differences between simulation and experiment using one faulty experimental example, and transfers knowledge from simulation to reality by addressing the transfer function effects. It addresses the transfer function by spectrum background estimation and minimum phase estimation while also selecting features that are invariant to the unmitigated effects of the transfer function. The new algorithm is demonstrated on simulated signals and measured transfer function, and on experimental signals with known fault sizes. The codes and the data of the study are available via the link: https://github.com/omriMatania/one_fault_shot_learning_for_gears_fault_severity_estimation.