Abstract:
Gearboxes are widely utilized as critical components in a large number of engineering applications. Gearboxes are prone to failures and therefore it is advantageous to utilise a condition-based maintenance (CBM) framework to infer the condition of its components. Various data-driven and physics-driven approaches have been developed for the CBM task. In this work, a hybrid approach is proposed where a data-driven and physics-driven approach are combined to infer the condition of the gearbox. The hybrid approach combines the advantages of both approaches and aims to overcome their respective limitations. For the physics-driven approach, a numerical gearbox model is developed. The modelling procedure for the numerical gearbox model introduced a novel approach to gear fault modelling which aims to generalize the introduction of gear faults to a simpler, unified framework. For the data-driven approach, a supervised convolutional neural network (CNN) is utilised to extract features from vibration signals and classify them simultaneously. By generating synthetic data from the physical model and feeding this to the CNN, a hybrid model is developed which may yield the potential for fault identification of the real asset. There is, however, no guarantee that the learned features from the synthetic data (source domain) are transferable to a new domain of signals (target domain), such as those from the real asset. Two transfer learning methods are utilised to calibrate the hybrid model for a change in input data. To investigate the efficacy of transfer learning calibration, two numerical experiments are constructed where the hybrid model is trained on perfect synthetic data (the source domain) and applied to noisy synthetic data with different vibration signatures (the target domain). The results show that an uncalibrated hybrid model fails to transfer to the target domain, but that the calibrated methods perform well on this transfer task. This work highlights the potential of transfer learning-calibrated hybrid methods for condition monitoring of gearboxes.