Meta deep learning based rotating machinery health prognostics toward few-shot prognostics

Abstract

Abstract Data-driven health prognostic is attracting more and more attention to machinery prognostic and health management. It enables machinery to realize predictive maintenance and rarely depends on prior knowledge of degradation mechanisms. However, cross-domain health prognostic may lack enough measured data as supports, and this bottleneck is particularly prominent in high-end manufacturing. As such, this paper aims to improve prediction performances under limited data coupled with variable working conditions. Meta learning is introduced into this field for the first time, and meta deep learning (MDL) based health prognostic methodologies toward few-shot prognostics are further proposed. To be specific, time–frequency images and time-series data are first picked up for abstracting domain-invariant degradation indicators based on the integration of covariance matrices and maximum mean discrepancy. Then the subtask and cross-subtask level gradient based optimization architecture is conducted to abstract more general degradation knowledge for prognostics models’ adaptation. Based on the architecture, two variants termed as meta convolutional neural network (meta CNN) and meta gated recurrent unit (meta GRU) are proposed to accomplish few-shot prognostics with different forms of degradation indicators. Thirdly, three cases of run-to-failed machinery experiments are employed for a large number of verifications to avoid unexpected results. Finally, appealing predictions compared with existing methods demonstrate the superiority of our proposed MDL health prognostics.