Machinery Probabilistic Few-Shot Prognostics Considering Prediction Uncertainty

Abstract

Few-shot learning based machinery prognostics are feasible for intelligent operation and maintenance with scarce monitoring data. In fact, single-point estimations of existing few-shot prognostics (FSP) algorithms suppress the high-reliability predictive maintenance. To alleviate this dilemma, the probabilistic few-shot prognostics problem is formulated to conquer the challenges of uncertainty estimation in FSP. We propose a novel Bayesian approximation enhanced probabilistic meta-learning (BA-PML) algorithm to convert learnable parameter uncertainty into final prediction uncertainty. It consists of two main components: the designed base probabilistic predictor and its corresponding episodic training strategy. The former reshapes Seq2Sep models with Bayesian theories and accomplishes variable-length degradation prediction. The latter follows the few-shot learning paradigm and extends variational inference driven posterior approximations to meta-level training that assists in mining general degradation knowledge from probabilistic aspects. Finally, run-to-failed vibration data proves our proposed BA-PML holds well-calibrated uncertainty prognostics under cross-domain decision-making tasks.