Meta-knowledge random attention update network for few-shot and anti-noise remaining useful life prediction

In industrial systems, the remaining useful life (RUL) prediction of industrial equipment is crucial to ensure system’s safe operation. Current RUL prediction models have made notable advancements, predominantly through the utilization of extensive degradation data exhibiting analogous patterns or approximate distributions. However, when the labeled degraded data is limited and the data is affected by noise, the distribution discrepancies between RUL data will increase, preventing these methods from effectively capturing shared knowledge among the data and struggling to obtain satisfactory prediction performance. In this respect, a new meta-knowledge random attention update network model is proposed for few-shot and anti-noise RUL prediction. First, we treat the learned kernel features as random latent variables in a Monte Carlo sampling manner. Then, the attention mechanism is introduced in the random kernel to realize the control of local degradation information and enhance the learning of specific knowledge by the model. In addition, to reduce the impact of unnecessary or noisy information on meta-knowledge, the integration of shared knowledge and specific information is implemented within the knowledge update procedure. Comprehensive experiments are performed on datasets pertaining to engine and bearing degradation to assess the efficacy of the proposed model, with the results confirming its superiority.